check-cep-vision

Import vision from check-cep-vision and call vision.clickByImage(page, 'assets/submit-btn.png', { region: 'main' }). The library takes a screenshot, runs OpenCV template matching, highlights the match, and clicks its center.

Use image-based locators when pages have dynamically generated CSS classes, canvas or WebGL content, or no accessible DOM elements. If a stable selector like data-testid or an ARIA role exists, DOM locators are faster and easier to maintain.

No, Lightpanda has no rendering pipeline so vision functions throw when BROWSER=lightpanda. Use vision.canScreenshot() to guard calls, or use hybrid functions like clickByImageOr which fall back to DOM selectors automatically.

The most common cause is a baseline mismatch: the template was captured at a different viewport, device scale factor, or outside the container with different font rendering. Re-capture the template inside the container at 1280x720 with DPR 1 and animations disabled.

Ambiguous means two candidates scored too close together. Narrow the search with a named region preset or custom rectangle so only one instance is in scope, or improve the template crop to include distinguishing context like labels.

Call vision.typeByImage with a template that includes the label plus input, and set clickOffset to shift the click from the label into the input field. It uses page.keyboard.type() for keystroke simulation; the hybrid typeByImageOr falls back to fill().